Show HN: Get conversational practice in over 20 languages by talking to an AI (talk.quazel.com)

787 points by Hadjimina ↗ HN
Hi everyone,

Let me introduce you to Quazel, where we want to enable people to talk their way to fluency.

We have all tried various language learning apps and tools, however, one aspect of language learning current services are really bad at is conversational practice. You might get a chat-like interface, but in the end, the conversation partner will only respond with a predefined "if the users say X I say Y".

With Quazel that's completely different. In completely dynamic and unscripted conversation you can talk about pretty much anything you want. For example, you can try ordering food at a restaurant and even hold a philosophical discussion with Socrates. Additionally, you can analyze the grammar of your responses or use hints to help you out when you get stuck.

We want to change how languages are learned from a grammar-centric approach to a more natural, conversation-focused one.

353 comments

[ 0.20 ms ] story [ 262 ms ] thread
It really looks cool. But, why it doesn't have English? Or meybe if it does have, why isn't showed first?

People who use language learning apps are focused in learning English most of the time. Still those who speak fluently are trying to improve their speaking and listening.

And the first language which appeared for me was Brazilian Portuguese which is my mother language.

[Edit] I was trying the Brazilian Portuguese mode, and apparently it is Portuguese from Portugal, because the answer given used a verb tense that is not common in Brazil.

(comment deleted)
Yeah. I been looking for a few months for services to improve my English in a cheap/free way. Until now some Discord channels have been my solution to practice conversation, but it would be nice to have this kind of service to speak anytime about any topic.
for what its worth what you just typed is completely indistinguishable from a native speaker, well done
The advantage of typing is that I can erase and improve what I'm trying to communicate. In real time I struggle thinking about pronunciation and if what I'm saying is correct or understandable by the other person.
Have you tried playing online multiplayer games? I found I was able to learn how to “think” in another language much better when I was under pressure to communicate something quickly in real time. It’s easy to fall into the trap of always translating to/from your native language which isn’t the end of the world when you’re reading or writing but it makes for pretty awkward and stilted conversations.
Try tandem to find people to talk with, it's free
To prevent DoS from a tonne of people using it just as AI conversation buddy when they know English perfectly well? Or said another way - I think there's lots of English AI conversation bots available, they're just not necessarily marketed as 'for learning English'.
You can't train AI using another AI. If you could, these people would just run the service locally.
I'm not sure I follow/see the link to my comment?
Hahah I have a super silly response for this one: It's simply because we have not gotten around to translate the UI and everything into other languages besides english. That's why we simple assume people know english, since they need to be able to know at least some english to use it. But it's high on our to-do list.
I think you misunderstood what they're missing: the ability to converse and practise English, given it's a foreign language for some people.
So then the UI would need to be in their non-English language, which is what they are saying.
Not necessarily. Learning enough English to understand the UI is much lower level than being able to converse.

It's also the difference between active and passive command of the language.

Also "right click -> Translate to <x>" takes care of the UI automatically (in Chrome at least) but still doesn't solve the issue that English can not used to converse with the AI.
> Not necessarily. Learning enough English to understand the UI is much lower level than being able to converse.

Exactly this. For example, I am learning German and have changed the language of my reading tablet's UI to German. That in itself is a useful learning tool.

My kids are struggling to learn English; would be very happy to use that with them. I can help in translating the UI in French if you'd like!
The UI can be in English for people who want to learn English. People who study languages are at different proficiency levels (very few people are blank slates with no knowledge of the language).
You don't need to translate English UI for training English because there are quite a lot of person don't talk / listen well even though he can write / read English quite okay.

It's because Reading / Writing can train oneself but conversation must be training in both sides, which always inadequate in some area, such as HK, since we don't talk in English but we can read quite well

(comment deleted)
This would be great for English. So many people need help with their conversational or small talk skills. Would be great for them to be able to practice even if they are native speakers. For public speaking, office communication training, self confidence development, etc.
Wow, I just tried the mandarin setting. Amazing work!
What's the tech stack? Whisper for language transcription, GPT-3 for NLP?
Isn't it generally expensive to build on top of GPT-3?
Depends on your workloads. 750 words will cost you 0.2 cents, that can be very expensive or reasonably priced, depending on what you do.
Cool!(but yeah english is really important!!)

which models do you use for speech recognition and text generation?

Yup english is top of at the top of the to-do list. We use some APIs from a few different big players. One that I want to highligh which is waayy better than the rest is Microsoft speech to text, which not only recognized what you say but can also do punctuation and capitalization. Instead of "do you like to eat pizza" which is what google speech to text would give you, microsoft recognizes it as "Do you like to eat pizza?". Small but important difference.
have you tried the new whisper model from openai?
Very nice! Please add arabic when you have a chance
Arabic is very high on the list but suuupppeeer tricky...
how is it worse than chinese or japanese ? i'm super curious..
Looks like there's still a few bugs, completely unable to have my voice recorded correctly or the install link to work on android
Yep we are super early and still have some bugs. Could you tell me what browser you were using? B.c. maybe you did not give the browser the permission to record stuff with the microphone and then things break.
Agreed that this is very cool, this is a massive unsolved niche in the language learning market. Execute effectively on this and you've got it made. So I noticed that on MacOS Firefox, the app never asked me for microphone permissions and thus didn't work at all. But worked well on Chrome (which I only have for testing purposes).
With the current state of AI text generation, isn't this a recipe for disaster? By talking to a chatbot while learning a language you are liable to pick up whatever (vast) errors and mistakes the bot makes.
That's a good point. However, I think we have reached a point where AI text generation has come pretty far and will only continue to get better. Even if you pick up one or two mistakes the A.I. makes but are otherwise fluent, isn't that already an improvement? (Not a rhetorical question, genuinely interested)
There's also the issue of the AI picking up your mistakes and adding more of its own to match the input. When AI Dungeon came out, I recommended it to a friend as a way to practice English, but when she showed me a playthrough, it turned out that the AI had adapted to her writing style full of misspellings and grammar errors, so it probably didn't help at all. (I haven't been able to test your chatbot yet to see whether it suffers from the same problem.)

In general, practicing something allows you to do it with less and less effort until you barely have to think about it at all, but if you're practicing making mistakes , that just means you'll be making mistakes without thinking about it. To get better, you also need to notice your mistakes to change what you're doing accordingly.

So I think a bot that can guess what you were trying to say and can tell you what you should've said instead might be a bit more useful.

IMHO, better practice with mistakes, then fix them later once you talk to native speakers, than not practice at all.

For kickstarting a language practice, this site honestly seems ideal..

Eh, I tested it with a language I speak and it was almost comically bad. Grammar constructions nobody would use, mixing vocabulary from two different variants in the same sentence, etc.
I don't want to say "recipe for disaster", but as someone working on a similar app, it's definitely a challenge. It's one of the reasons I stayed away from a pure ML/language modelling approach. It's really easy for models to get things wrong - for example, elsewhere in this thread there's a Chinese sentence about arranging furniture but the word "arrange" isn't really right (it's more "arrange" in the sense of "arranging a delivery", rather than "arranging a number of objects", and sounds like it's been auto-translated from English).

When I started I wanted to be sure that every single word/sentence was "correct" (i.e. a native speaker agreed that all the words/grammar were correct/natural in that context). LMs have improved since I started, but for me, that requirement still stands and ML approaches still need a heavy human touch IMO.

I recognize the voice of the Dutch examples from Duolingo. What model are you using to generate these?
From the FAQ:

> It might, however, also be that the conversation partner is feeling uncomfortable talking about certain things.

Does this mean that users talk to actual people or am I not understanding this correctly?

So in the background we are checking if the stuff you are talking about is "risky". If that is detected, the conversation partner will just answer with something like "hmmm I don't feel comfortable talking about that" to keep it all in safe territory.
How do you scale that? Or is the goal to automate that as well with AI?
Do you really mean risky, or do you mean risqué?
I am stuck at the loading screen when selecting Dutch. BTW I am using Firefox
Nice. Please add Icelandic some time soon. I'd probably be willing to pay for that.
Very cool. Just tried it - works really well. Are you raising money? How can I invest?
if you are serious about this (which would be dope), send an email to philipp@quazel.com
Given that a lot of conversational chat bots end up devolving into... lets charitably call it 'repugnant ideas'... I am not sure that learning to talk from them is wise.

Interesting concept, but I'm going to pass for now. I don't want to be talking to a friend and find that I accidently espoused the ideas of the third reich.

According to the author, they censor various topics
Given that's it Azure, that's probably done for them whether they want to or not.
Could you add Icelandic ? It’s spoken by just over 300,000 people and I consider it be “endangered”
Very cool idea, but doesn't it have a level adjustment? For example, I'm learning Chinese but my level is quite basic at the moment, it would be cool if the bot could adapt its vocabulary.
Yep, this is a highly requested feature..but as you can imagine a bit tricky. We are working on it. However, I am using it for Greek and I'm on an early A2 level...so I think if you are at that point you can probably get pretty far by using the hints and translator.
Yes, to be honest even at my level (also A2) it's pretty useful, thanks to the integrated translator. And it already kind of adapts by default, in the sense that if I talk about easy topics (daily schedule, and things like that) it will talk about easy topics as well, so it's not like it's coming up with quantum mechanic terms.

It's a suggestion for improvement but it's already very good and I think I'm going to use it a lot.

By the way, just in case you don't know and you find it interesting, in the case of Chinese there was a crowdfunding campaign about a hardware speaker that would do something like this (https://webmarketkings.com/lily-chinese-scam/). It never delivered. I was one of the backers... but well, that's life with Kickstarter and the like, sometimes.

您好!我也正在学中文,但是我也很初学者。What do you do for speaking practice? I live in China, so I get plenty of listening practice, but not too much on speaking (it's hard to keep people's attention with broken sentences :P). I made this an (https://letsyiya.com) to help with that. It's not as good as I'd like it to be, but I've found it helpful. I'd be glad to hear your setup :)
你好! 我住在西班牙,所以我没有太多机会说。There is an academy near me but it's hard to find time (busy job) so my listening practice is mostly apps (Chairman's Bao and an Anki deck with audio) and videos sometimes, and speaking practice is almost none (repeating Anki flashcards and sometimes speaking to Google Translate to see if it comes up with what I'm trying to say... this website is better in that respect!)

BTW, your link doesn't work for me right now.

Great idea, but when trying french the UI hangs indefinitely waiting for my input (spinning wheel type of waiting sign), after clicking on microphone (latest Firefox with uBlock origin)
Loading screen doesn't go away in Firefox after choosing a language.
hmm strange just tried it. What langauge are you using?
This happened to me too. I'm using Brave and I needed to disable its Shields. Maybe you have some extension impacting the site, like an ad blocker?
I saw that bug too, if you refresh it enough, it goes away.
Hey this is a cool idea. I tried it with Russian -- on Chrome at least, nothing works with my mic (but the mic permission worked, it just doesn't "listen" or show feedback when I speak).

Hey, hit me up by email (see my profile), I'd love to talk to you and help out with anything you might need. Languages are one of my passions; I'm a polyglot (fr/el/sv/ru/en) and have a lot of opinions on language learning and teaching.

This looks so nice.

On a side note, I used Duolingo for about 7 or 8 months (I spent like 15-20 minutes) and learned a little Japanese and Portuguese (European). I think Duolingo is an enjoyable app. Then I lost interest :( I started again. Let us hope this time I will finish all levels.

Any chance of getting Japanese and Portuguese (European)?

(comment deleted)
Here’s what I think would be a killer feature. Have it use AI to adapt the conversation level to the level of learner. Smaller words, repetition, slower.

If you notice this is how adults talk to babies and children to help them learn the language.

I'd agree but it depends on what AI engine that he uses to generate the conversation
There are many possible knobs on such a tool that could be wonderfully useful. I'd love to press a button that provides critical feedback: marking for errors or suggesting improvements for style (grammar, better word choice, useful idioms, etc).

It'd also be great if my sentence was scored for 'natural-ness' -- how close was it to native speech? If this could be done regionally, that'd be better still (e.g. Spanish as spoken in Spain vs Mexico vs Argentina, etc).

I'd also love to know what age level am I speaking at. A child? A teen? An adult? A more educated person? A professional author? Am I being ambiguous? Too direct?

What is the tone I am using? Friendly? Demanding? Too formal? Culturally clueless? Improper to someone older or to the opposite sex?

In fact, these same feedbacks could be really valuable even in your own native language. Learning how to communicate more clearly, precisely, positively, professionally, respectfully, literately, or even in a different style would be great practice to improve a person's basic communication skills in any language.

We also see a lot of potential in giving the most precise feedback possible. In that way each individual learner can benefit the most. The development is currently on-going and we are looking forward to the next steps!
I really like this. Especially with the repetitions and automatically having the TTS be slower for beginners.
Absolutely. Would also be great to allow the user to select the conversation level in the event the AI doesn’t detect the desired one or one wants to try to converse at a more difficult or less difficult level for practice.

Some languages have well defined language levels for foreigners learning the language, such as HSK levels for Chinese. Would be cool to leverage that somehow.

This is really great and well executed. Well done!

One problem I'm beginning to notice is that the connection to the realtime transcription is failing (I just see a loading screen when I click to reply). Maybe it's due to heavy load?

I notice you are using Azure Cognitive Services for the transcription at the moment. Out of curiosity, did you consider any other services for this? (I'm building a transcription-based app myself and I'm worrying about the ability to handle lots of connections at once)